TV → LTX Adapter 🎬

Of the three adapters in the TensorVizion Director pack, this is the one that feels closest to magic, because LTX is the only backend that lets you render a whole timeline as one generation. TV → LTX Adapter 🎬 takes the shot list from TV Director and flattens it into a single positive/negative conditioning pair plus one guide-latent stack - no per-shot loops, no stitching, no sampler gymnastics. You sample once, and the entire timeline comes out as a continuous clip.

That's only possible because of how LTX actually works. Unlike Wan or Hunyuan, which condition per-clip, LTX supports injecting reference frames at arbitrary positions mid-generation through its guide-latent mechanism (LTXVAddGuide). Lightricks never really advertised it - the community had to discover that LTX could splice frames and video fragments into a generation all along, and LTX Director got built as a wrapper around exactly that. This adapter is doing the same trick for a generic timeline: it doesn't reimplement any LTX math, it calls ComfyUI-LTXVideo's own LTXVConditioning and LTXVAddGuide node classes directly.

How it works

For each shot in the timeline, the adapter text-encodes the shot's prompt and negative, runs them through LTXVConditioning with your timeline's fps, and merges them into a running positive and negative. Then it resolves the shot's reference image (if it has one) and injects it as a guide frame at the right spot - first pins it to the shot's start, last to its end, middle/reference to the midpoint - at the shot's per-shot strength. If no shot has an image, you get a plain empty T2V latent sized from the timeline's global width/height.

Inputs and outputs that matter

Three required inputs: timeline (the DIRECTOR_TIMELINE from TV Director), clip (your LTX text-encoder CLIP), and vae. Nothing optional - a T2V timeline needs exactly these.

Outputs, all four of which you'll use: positive and negative (CONDITIONING) go into an LTX sampler, guide_latent (LATENT) is the frame stack with your references baked in, and num_frames (INT) tells you the total length - feed it to LTXVBaseSampler or a KSampler set up for your LTX checkpoint, then decode and save.

Install

The pack itself clones like any custom node, but this adapter needs the backend:

cd ComfyUI/custom_nodes/
git clone https://github.com/TensorVizion/ComfyUI-Director-Node-Pack tensorvizion-director

then install ComfyUI-LTXVideo via ComfyUI Manager. The adapter checks for it at execution time and raises an error naming the exact missing nodes - it won't silently produce garbage.

Honest notes

This is where LTX's speed shines: a full shot list in a single generation, which suits rapid iteration (its quality ceiling still trails Wan 2.2, but you can do ten of these in the time one Wan render takes). Two caveats: transitions like crossfade are still intent-only, so shots concatenate rather than blend, and the adapter targets the node APIs as of mid-2026 - if Lightricks changes LTXVAddGuide's signature upstream, this needs a matching update. It's also a young pack with no community track record, so test on a two-shot timeline before committing to a long one.